StreamDiffusion SDXL → Scope (LongLive) via VACE Depth

I’ve been experimenting with what “long-form” Scope generation could look like when seeded from StreamDiffusion V1.

The workflow was:

• Create a Feedback Trail visual effect in TouchDesigner
• Generate “a trail of flowers” in TouchDesigner StreamDiffusion SDXL with the visual
• Send that output into TDDepthAnything in TouchDesigner
• Use the resulting depth map as input for Scope LongLive via VACE Depth, where I prompted the idea of “fiery, blooming flowers” using a LongLive-style prompt (with help from ChatGPT based on the Scope documentation)

The results were stronger than I expected.

Compared to running Scope alone, the video feels more cinematic, structured, and spatial. By letting StreamDiffusion establish a visual world first, Scope has a much clearer foundation to build on instead of inventing everything from scratch.

What stood out to me:

• Improved temporal consistency
• Much less jitter and flashing than typical StreamDiffusion V1 outputs
• You can switch to other prompts, and Scope will adapt well to different them, rather than getting “stuck” in one style
• Higher quality output with less noise blur.

This suggests StreamDiffusion can work well as a world-building layer, while Scope + VACE functions as a cinematic refinement layer that evolves the scene over time.

Side-by-Side SDXL vs LongLive comparision 

I also ran a side by side benchmark (thanks for the idea @oceanradiostation) where I repeated the same process entirely within StreamDiffusion, using a first StreamDiffusion V1 instance to generate visuals and then feeding that output into a second StreamDiffusion instance, so I could directly compare its refinement and consistency against the StreamDiffusion to Scope LongLive with VACE Depth pipeline.

Prompt Adherance and Prompt Travel

What is really cool is that the outputs of Scope LongLive stay anchored to the concept. For example, if you prompt something like underwater flowers, Scope produces a cohesive seafloor environment with bubbly, blooming roses rather than just making the flowers more blue and the background a little bit bubbly. That level of thematic consistency really surprised me in a good way.

Ontop of that, changing prompts while the model is running is pretty cool. There's an option to change the transition length as well. The prompt transitions are not perfect, sometimes it freezes up, but perhaps that's because of my shoddy wifi. Overall, in my opinion the travel is noticeably better than the StreamDiffusion V1 cloud version.

In conclusion, the side-by-side tests suggest there are useful cases for both approaches. However, if you have a very specific idea that you want to realize from a prompt, it may be worth seriously considering Scope. It seems to hold onto prompts more faithfully over time, which is the most important difference to me.

Curious to hear what other people think. If you want to try for yourself, I've left the depth map as a file you can download, as well as the prompt.